Synthetic training data glossary
Updated 2026-10-12
- Synthetic datasource
- Data produced by a modeled or simulated process instead of direct observation. Here the focus is imagery and sensor observations used in vision workflows.
- Domain randomizationsource
- Varying scene properties such as lighting, color and object position across generated samples, so a dataset contains more than one fixed appearance.
- Ground truthsource
- Reference labels or scene properties accompanying an observation, such as object identities and segmentations. Generated labels still need to match the task’s definitions.
- Semantic segmentationsource
- Assigning pixels to a class, such as a class of object. It describes category membership rather than distinguishing every individual object.
- Instance segmentationsource
- Separating individual object instances in an image. Two objects of the same class can have different instance masks.
- Depth mapsource
- An image-like representation of scene depth for each pixel. Confirm the generator’s units and distance convention before using it as a target.
- Optical flowsource
- An image motion field describing the movement of visible points between frames, useful for learning or evaluating motion estimation.
- Annotation exportsource
- The handoff of labels in a format the next tool can read. BlenderProc, for example, documents COCO and BOP annotation writers alongside rendered observations.